Auto mode classifier persistently unavailable — 2+ weeks, reproduces across accounts, machines, OSes, and CLI versions

Auto mode has been unable to approve any Bash/Write/Edit action for ~2 weeks. Every attempt returns:

claude-{model} is temporarily unavailable, so auto mode cannot determine the safety of {tool} right now.

Systematically ruled out the usual causes:

| Variable changed | Result |
|---|---|
| Session model: Opus 5 → Sonnet 5 | Same failure, error message just swaps the model name |
| dangerouslyDisableSandbox: true on the Bash call | Still routed through classifier, still fails |
| In-repo, fully-allowlisted paths vs. out-of-repo paths | Elevated-risk actions (external paths, self-permission-file edits) fail; plain allowlisted in-repo commands succeed by never hitting the classifier at all |
| CLI version: 2.1.220 → 2.1.221 (claude update) | Same failure before and after |
| Machine: Linux (native install) → 2 separate Windows machines | Same failure on all three |
| Account: tested with 2 different accounts | Same failure on both |

Only acceptEdits/default permission mode works, since it skips the classifier entirely. This has been reproducible on every attempt for two weeks straight — not intermittent, not overload-shaped.

Given it survives a device change but not an account change, this looks like an account/org-scoped classifier assignment (per #67542's "no fallback model" mechanism, possibly manifesting as disguised rate-limiting per #60438) rather than a general infra outage.

Related: #67542, #60438, #69726, #68437, #63819